An object-oriented center-oriented semantic-aware three-dimensional gaussian scene reconstruction method and system
By distinguishing between foreground and background through a semantic awareness mechanism, the 3D Gaussian scene reconstruction process is optimized, solving the problem of insufficient foreground detail reconstruction under sparse views and achieving higher reconstruction accuracy and robustness.
Patent Information
- Application Number
- CN202610514055.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-14
AI Technical Summary
Under sparse view conditions, existing technologies have limited effect on the detailed reconstruction of foreground objects by 3D Gaussian sputtering models, which cannot effectively distinguish between foreground and background, leading to geometric distortion and overfitting problems.
By introducing a semantic awareness mechanism, pixel-level semantic information is provided using 2D segmentation masks to distinguish between foreground and background. Differentiated Gaussian point densification and weighted rendering error calculation are designed to optimize the 3D model generation process.
It significantly improves the accuracy and rendering quality of foreground object reconstruction under sparse view conditions, reduces the risk of overfitting, and enhances the applicability and robustness of 3D models.
Smart Images
Figure CN122391491A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of 3D scene reconstruction technology, and in particular to an object-centered semantic-aware 3D Gaussian scene reconstruction method and system. Background Technology
[0002] With the increasing demand for real-time, highly realistic rendering, 3D Gaussian sputtering (3DGS) has become a key technology in computer vision. However, 3DGS is prone to geometric distortion due to overfitting when the input viewpoint is sparse, a challenge that has attracted widespread research attention. This content-agnostic stochastic strategy treats all scene regions equally during the optimization process, limiting the model's ability to represent details of key foreground objects. To address this difficulty, Park et al. proposed DropGaussian, a concise and effective structure regularization method. It is a Gaussian (normal) distribution-based regularization method in deep learning, a variant of Dropout. Its core principle is to introduce the randomness of the Gaussian distribution to smoothly adjust the activation values or weights of neurons in the neural network, replacing the hard truncation (randomly setting to 0) of traditional Dropout, thus more gently preventing model overfitting. This method randomly removes some Gaussian points during training, allowing the model to learn a more reliable scene structure, thereby improving performance in sparse view scenarios.
[0003] The success of DropGaussian demonstrates that optimizing scene models by altering the state of Gaussian points is feasible. However, its random dropout strategy has a problem: it is indiscriminate. It treats all Gaussian points in the scene equally, regardless of whether they belong to foreground objects requiring detailed rendering or to relatively minor background elements. While this approach effectively prevents model errors, its effect on improving the sharpness of foreground object details is limited. This issue is critical in many applications, such as creating high-precision digital humans, editable game models, and AR interactions, because these applications require foreground details to be far sharper than the background. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this application provides an object-centered semantic-aware 3D Gaussian scene reconstruction method and system. To achieve the above objectives, the technical solution of this application is as follows: A semantically perceptive 3D Gaussian scene reconstruction method centered on objects, comprising the following steps: processing the input image to generate an initial Gaussian point cloud; providing pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generating a two-dimensional segmentation mask, on which a mask value is defined based on the recognition results of the foreground and background; projecting the center of each Gaussian point in the initial Gaussian point cloud onto the two-dimensional segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point, and determining the semantic label of each Gaussian point based on the mask value at the corresponding 2D position; performing differentiated Gaussian point densification for the foreground and background; defining a weight function based on the foreground mask to construct a weight map, with the foreground weight being greater than the background weight, calculating the weighted rendering error based on the constructed weight map, and generating a 3D model; and rendering a new view based on the generated 3D model. Furthermore, a motion recovery structure is used to process the input sparse view image to generate a point cloud composed of sparse 3D points, with each 3D Gaussian point set based on the point cloud. Furthermore, a segmentation model is used to perform 3D recognition on each 2D image, distinguishing between the foreground and background regions, and providing a corresponding 2D object segmentation mask. The 2D segmentation mask is a binary image, where a region with a pixel value of 1 represents the foreground region and a region with a pixel value of 0 represents the background region. Furthermore, temporary semantic labels are generated based on the two-dimensional segmentation mask of each view. By combining the temporary semantic labels under all visible views, the semantic label of each Gaussian point is determined. Furthermore, foreground and background thresholds are established, with the background threshold being greater than the foreground threshold. For Gaussian points whose semantic labels are defined as foreground regions, densification is triggered when the gradient norm of their loss function is greater than the foreground threshold. For Gaussian points whose semantic labels are defined as background regions, densification is triggered when the gradient norm of their loss function is greater than the background threshold. This application further provides an object-centered semantic-aware 3D Gaussian scene reconstruction system, comprising: an initial Gaussian point cloud generation module for processing input images to generate an initial Gaussian point cloud; a 2D segmentation module equipped with a segmentation model, wherein the segmentation model provides pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generates a 2D segmentation mask, wherein a mask value is defined on the 2D segmentation mask based on the recognition results of the foreground and background; a semantic labeling module for projecting the center of each Gaussian point in the initial Gaussian point cloud onto the 2D segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point, and determining the semantic label of each Gaussian point based on the mask value at the corresponding 2D position; a Gaussian point densification module for performing differentiated Gaussian point densification for the foreground and background; a loss calculation module for defining a weight function based on the foreground mask to construct a weight map, wherein the foreground weight is greater than the background weight, calculating the weighted rendering error based on the constructed weight map, and generating a 3D model; and an output module for rendering and outputting the generated 3D model. Furthermore, the initial Gaussian point cloud generation module uses a motion recovery structure to process the input sparse view image to generate a point cloud composed of sparse three-dimensional points, and each 3D Gaussian point is set based on the point cloud. Furthermore, the two-dimensional segmentation module performs three-dimensional recognition on each two-dimensional image through a segmentation model, distinguishes between the foreground and background regions, and provides a corresponding two-dimensional segmentation mask for the object. The two-dimensional segmentation mask is a binary image, where the region with a pixel value of 1 represents the foreground region and the region with a pixel value of 0 represents the background region. Furthermore, the semantic labeling module generates temporary semantic labels based on the two-dimensional segmentation mask of each view, and combines the temporary semantic labels under all visible views to determine the semantic label of each Gaussian point. Furthermore, the Gaussian point densification module sets a foreground threshold and a background threshold, with the background threshold being greater than the foreground threshold. For Gaussian points whose semantic labels are defined as foreground regions, densification is triggered when the gradient norm of their loss function is greater than the foreground threshold. For Gaussian points whose semantic labels are defined as background regions, densification is triggered when the gradient norm of their loss function is greater than the background threshold. The beneficial effects of this application are as follows: by introducing semantic information and utilizing the semantic priors provided by the 2D segmentation model to integrate into the optimization process of 3D Gaussian scene reconstruction, a shift from random processing to semantic guidance is achieved. Specifically, by semantically perceiving and distinguishing foreground objects from background regions and generating semantic labels, differentiated processing is implemented for Gaussian points: the Gaussian point growth rule is modified to allow Gaussian points in the foreground region to split more easily to generate more details; a weighted loss function is designed to give higher weight to the foreground region when calculating rendering errors. This differentiated optimization mechanism effectively solves the problems of geometric distortion and insufficient foreground detail fidelity caused by random and indiscriminate strategies in existing technologies, especially reducing the risk of overfitting under sparse view input, and improving the accuracy and rendering quality of 3D models. Furthermore, experimental verification on multiple benchmark datasets shows that this method, while maintaining robustness, significantly outperforms existing methods in terms of object-level reconstruction metrics, enhancing the applicability of scene reconstruction and making it suitable for application scenarios requiring high-precision foreground details. Attached Figure Description
[0005] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.
[0006] Figure 1 This is a flowchart illustrating an embodiment of this application; Figure 2 The rendering result of the Blender dataset obtained through the embodiments of this application in a sparse view. Detailed Implementation
[0007] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.
[0008] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit this application. The terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing description of the drawings of this application are intended to cover non-exclusive inclusion.
[0009] Based on this, this application provides an object-centered semantic-aware 3D Gaussian scene reconstruction method. This application reconstructs a 3D Gaussian scene based on several input 2D images. The input images are several 2D images from different perspectives, and can be sparse. To improve the 3D Gaussian scene reconstruction under sparse perspectives, such as... Figure 1 As shown, the present invention includes a preprocessing stage, a semantically guided structure optimization stage, and a loss function calculation stage.
[0010] The preprocessing stage aims to recover a coarse 3D geometry of the scene from a limited set of input images (i.e., sparse views), providing pixel-level semantic information for each input training image to distinguish foreground objects and background regions within the scene. Specifically, it includes the following steps: 1.1. Process the input two-dimensional images to generate an initial Gaussian point cloud: 1.2. Provide pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generate a two-dimensional segmentation mask. Define mask values on the two-dimensional segmentation mask based on the recognition results of foreground and background.
[0011] Specifically, in some embodiments of this application, step 1.1 uses Structure from Motion (SfM) technology to generate an initial Gaussian point cloud. SfM is a classic multi-view geometry algorithm that automatically estimates camera parameters (such as position and orientation) for each image by analyzing the feature point matching relationships in a set of two-dimensional images. The initial point cloud generated by SfM is the starting point for 3DGS model optimization, determining the basic spatial structure of the scene, and its direct output is a point cloud composed of sparse three-dimensional points. These sparse points are used as seed points for initializing the 3D Gaussian Splatting (3DGS) model. Each 3D Gaussian point (G... i The initial position, color, and other attributes of the point cloud will be set based on this point cloud. The sparse point cloud generated by the motion recovery structure has a clear physical spatial correspondence, providing a stable spatial distribution basis for Gaussian points.
[0012] Specifically, in some embodiments of this application, step 1.2 uses a pre-trained segmentation model for each training view. Provide foreground mask This pre-trained segmentation model can identify foreground objects and background regions in a 2D image and generate corresponding binary images, or 2D segmentation masks, based on the identified foreground objects and background regions. Regions with a pixel value of 1 represent identified foreground objects (such as people or vehicles), while regions with a pixel value of 0 represent the background. These masks serve as valuable semantic prior knowledge, introducing semantic information from the 2D image space into the 3D optimization process. This directly guides the subsequent differential optimization process of 3D Gaussian points, enabling the system to "understand" which parts of the scene are the key objects requiring more refined reconstruction.
[0013] The input two-dimensional images are preprocessed in a preprocessing stage to obtain an initial Gaussian point cloud and a two-dimensional segmentation mask. The initial Gaussian point cloud is a point cloud composed of several 3D Gaussian points representing the rough three-dimensional geometric structure of the identified scene. The two-dimensional segmentation mask contains pixel-level semantic information that distinguishes foreground objects and background regions in the scene.
[0014] In this context, the semantic-guided structure optimization stage is each Gaussian point. Assign a semantic label (where 1 represents the foreground), and based on this semantic label, differentiated Gaussian point densification is achieved. Specifically, this includes the following steps: 2.1. Project the center of each Gaussian point in the initial Gaussian point cloud onto the 2D segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point, and determine the semantic label of each Gaussian point in the view based on the mask value at the corresponding 2D position. 2.2. Differentiated Gaussian point densification for foreground and background; Specifically, in some embodiments of this application, step 2.1 involves using the Gaussian point... Projecting onto this view yields the 2D position. The label under this view is determined by the mask value. Decide: .
[0015] Due to potential errors in segmentation from a single viewpoint, the final label of Gaussian points... The semantic label can be determined by a vote of the label from all visible views. Specifically, for each Gaussian point, the following operations are performed: project its center coordinates onto all visible views; if the projected point falls into the foreground mask area in a view, the foreground vote is incremented by 1; otherwise, the background vote is incremented by 1; finally, the semantic label is determined according to the majority voting principle.
[0016] Specifically, in some embodiments of this application, step 2.2 no longer uses a globally uniform gradient threshold. Instead, it is replaced with two separate thresholds: one for the foreground. And a background The condition for whether a Gaussian point needs to be compacted. The definition is as follows: ; By setting This lowers the threshold for new details to grow in the foreground region, thus allowing more model capacity to be used to characterize the foreground. The differential thresholding mechanism achieves adaptive resource allocation at the algorithm level, avoiding excessive densification of the background region and consuming computational resources, while ensuring that the foreground region quickly forms a high-precision geometric representation in the early optimization stage.
[0017] The semantic-guided structure optimization stage utilizes semantic information to achieve differentiated and targeted optimization of Gaussian point clouds, making up for the shortcomings of traditional methods where the random strategy content is unknown. This simple, semantically guided operation can effectively focus the model's expressive power on key areas in the scene.
[0018] The loss function calculation stage is used to calculate the loss function, which quantifies the difference between the "RGB image rendered by 3D Gaussian" and the "real training view image," thereby driving the model to iteratively optimize towards a better direction. Specifically, it includes the following steps: 3.1. Define a weight function based on the foreground mask to construct a weight graph, where the foreground weight is greater than the background weight; 3.2. Calculate the weighted rendering error based on the constructed weight graph and generate a 3D model; Specifically, in some embodiments of this application, the loss function calculates the focus-weighted photometric loss, and step 3.1 utilizes the foreground mask. Construct a spatially variable weight graph : ; in, It is a hyperparameter.
[0019] Specifically, in some embodiments of this application, step 3.2 uses focus-weighted L1 loss. The calculation is as follows: ; Where N is the total number of pixels in the image. This weighting amplifies the rendering error of the foreground region when calculating the total loss, resulting in a larger gradient signal that forces the optimizer to prioritize correcting the color and brightness of the foreground.
[0020] The object-oriented semantic-aware 3D Gaussian scene reconstruction method provided in this application embodiment can be applied to terminal devices. In this case, the terminal device is the execution subject of the object-oriented semantic-aware 3D Gaussian scene reconstruction method provided in this application embodiment. This application embodiment does not impose any restrictions on the specific type of terminal device.
[0021] In some examples, the terminal device can be any terminal device such as a mobile phone, tablet computer, laptop computer, smart TV (e.g., smart screen), ultra-mobile personal computer (UMPC), handheld computer, netbook, personal digital assistant (PDA), wearable electronic device, in-vehicle device, virtual reality device, etc.
[0022] Corresponding to the object-oriented semantic-aware 3D Gaussian scene reconstruction method described in the above embodiments, this application also provides an object-oriented semantic-aware 3D Gaussian scene reconstruction system. Each module of this system can implement each step of the object-oriented semantic-aware 3D Gaussian scene reconstruction method. The system includes: The input module is used to input images; The initial Gaussian point cloud generation module is used to process the input image and generate an initial Gaussian point cloud; The two-dimensional segmentation module is equipped with a segmentation model. The segmentation model provides pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generates a two-dimensional segmentation mask. The mask value is defined on the two-dimensional segmentation mask according to the recognition results of the foreground and background. The semantic labeling module projects the center of each Gaussian point in the initial Gaussian point cloud onto the two-dimensional segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point. Based on the mask value at the corresponding 2D position, the semantic label of each Gaussian point in that view is determined. The Gaussian point densification module is used to perform differentiated Gaussian point densification on the foreground and background. The loss calculation module defines a weight function based on the foreground mask to construct a weight map, with the foreground weight being greater than the background weight. It calculates the weighted rendering error based on the constructed weight map and generates a 3D model. The output module renders and outputs the generated 3D model.
[0023] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0024] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0025] To evaluate the object-oriented, semantically aware 3D Gaussian scene reconstruction method and system provided in the application embodiments, three representative benchmark datasets were used: LLFF, Mip-NeRF 360, and Blender. The following validation process follows the settings used above, employing the same partitioning for the LLFF, Mip-NeRF 360, and Blender datasets, containing 3, 12, and 8 training views, respectively. For LLFF and Mip-NeRF 360, the downsampling rate was set to 8, while for Blender, a downsampling rate of 2 was used.
[0026] For quantitative evaluation, this application uses three metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Patch Similarity (LPIPS), which are widely used in this field. PSNR measures the average peak error between a rendered image and a real image. SSIM calculates structural similarity based on brightness, contrast, and texture information. LPIPS calculates perceptual distance by utilizing learned features, which is useful for identifying potential differences that traditional metrics fail to reflect.
[0027] Performance Evaluation: The method of this application embodiment was evaluated under the LLFF dataset setting. Table 1 shows the quantitative comparison results with state-of-the-art (SOTA) methods. The method of this application embodiment outperforms existing NeRF-based and 3DGS-based methods on all metrics. SA-3DGS achieves a PSNR of 21.13. The method of this application embodiment also achieves SOTA performance on SSIM and LPIPS metrics, with scores of 0.740 and 0.179, respectively.
[0028] Table 1. Performance comparison of sparse view synthesis on the LLFF dataset The best, second-best, and third-best results are marked with different shades of gray.
[0029] SA-3DGS was further evaluated using a 12-view sparse setting on the Mip-NeRF360 dataset. As shown in Table 2, the method of this embodiment exhibits competitive performance across all metrics. Specifically, SA-3DGS achieved the highest scores on both PSNR and SSIM, reaching 19.90 and 0.580, respectively.
[0030] Table 2 Comparison of sparse view synthesis performance on the Mip-NeRF360 dataset The best, second-best, and third-best results are marked with different shades of gray.
[0031] Finally, SA-3DGS was evaluated on the Blender dataset. Figure 2 As shown, the method in this embodiment exhibits competitive performance across all metrics. Specifically, SA-3DGS achieved the highest score of 25.48 on the PSNR metric.
[0032] In summary, this application introduces a simple yet powerful method—the object-centric semantic-aware 3D Gaussian scene reconstruction method SA-3DGS—aimed at solving the foreground detail reconstruction problem in 3DGS under sparse view settings. The core idea of this method is to elevate the random regularization of DropGaussian to a semantically guided deterministic optimization process. By utilizing 2D segmentation masks, a differentiated optimization mechanism is designed, enabling Gaussian points belonging to foreground objects to achieve stronger structural refinement capabilities.
[0033] This simple, semantically guided operation effectively focuses the model's expressive power on key areas of the scene. Therefore, even under the extremely challenging sparse view conditions, SA-3DGS can significantly improve the reconstruction fidelity of foreground objects while maintaining strong geometric robustness.
[0034] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A semantically perceptive 3D Gaussian scene reconstruction method centered on objects, characterized in that, Includes the following steps: The input image is processed to generate an initial Gaussian point cloud; Provide pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generate a two-dimensional segmentation mask. The mask value is defined on the two-dimensional segmentation mask based on the recognition results of the foreground and background. Project the center of each Gaussian point in the initial Gaussian point cloud onto the 2D segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point, and determine the semantic label of each Gaussian point based on the mask value at the corresponding 2D position. Differentiated Gaussian point densification is applied to the foreground and background. A weighting function is defined based on the foreground mask to construct a weight map, with the foreground weight being greater than the background weight. The weighted rendering error is calculated based on the constructed weight map to generate a 3D model. Render a new view based on the generated 3D model.
2. The object-centered semantic-aware 3D Gaussian scene reconstruction method according to claim 1, characterized in that, The input sparse view image is processed using a motion recovery structure to generate a point cloud composed of sparse 3D points, with each 3D Gaussian point set based on the point cloud.
3. The object-centered semantic-aware 3D Gaussian scene reconstruction method according to claim 1, characterized in that, The segmentation model performs 3D recognition on each 2D image, distinguishes between the foreground and background regions, and provides a corresponding 2D segmentation mask for the object. The 2D segmentation mask is a binary image, where the region with a pixel value of 1 represents the foreground region and the region with a pixel value of 0 represents the background region.
4. The object-centered semantic-aware 3D Gaussian scene reconstruction method according to claim 1, characterized in that, Temporary semantic labels are generated based on the 2D segmentation mask of each view. By combining the temporary semantic labels from all visible views, the semantic label of each Gaussian point is determined.
5. The object-centered semantic-aware 3D Gaussian scene reconstruction method according to claim 1, characterized in that, Set foreground and background thresholds, with the background threshold being greater than the foreground threshold; For a Gaussian point whose semantic label is defined as a foreground region, densification is triggered when the gradient norm of its loss function is greater than the foreground threshold. For Gaussian points whose semantic labels are defined as background regions, densification is triggered when the gradient norm of their loss function is greater than the background threshold.
6. An object-centered, semantically perceptive 3D Gaussian scene reconstruction system, characterized in that, include: The initial Gaussian point cloud generation module is used to process the input image and generate an initial Gaussian point cloud; The two-dimensional segmentation module is equipped with a segmentation model. The segmentation model provides pixel-level semantic information for each input image to distinguish foreground objects and background regions in the scene, and generates a two-dimensional segmentation mask. The mask value is defined on the two-dimensional segmentation mask according to the recognition results of the foreground and background. The semantic labeling module projects the center of each Gaussian point in the initial Gaussian point cloud onto the two-dimensional segmentation mask of each view to obtain the corresponding 2D position of each Gaussian point, and determines the semantic label of each Gaussian point based on the mask value at the corresponding 2D position. The Gaussian point densification module is used to perform differentiated Gaussian point densification on the foreground and background. The loss calculation module defines a weight function based on the foreground mask to construct a weight map, with the foreground weight being greater than the background weight. It calculates the weighted rendering error based on the constructed weight map and generates a 3D model. The output module renders and outputs the generated 3D model.
7. The object-centered semantic-aware 3D Gaussian scene reconstruction system according to claim 6, characterized in that, The initial Gaussian point cloud generation module uses a motion recovery structure to process the input sparse view image and generate a point cloud composed of sparse three-dimensional points. Each 3D Gaussian point is set based on the point cloud.
8. The object-centered semantic-aware 3D Gaussian scene reconstruction system according to claim 6, characterized in that, The two-dimensional segmentation module performs three-dimensional recognition on each two-dimensional image through a segmentation model, distinguishes between the foreground and background regions, and provides a corresponding two-dimensional segmentation mask for the object. The two-dimensional segmentation mask is a binary image, where the region with a pixel value of 1 represents the foreground region and the region with a pixel value of 0 represents the background region.
9. The object-centered semantic-aware 3D Gaussian scene reconstruction system according to claim 6, characterized in that, The semantic labeling module generates temporary semantic labels based on the two-dimensional segmentation mask of each view, and determines the semantic label of each Gaussian point by combining the temporary semantic labels under all visible views.
10. The object-centered semantic-aware 3D Gaussian scene reconstruction system according to claim 6, characterized in that, The Gaussian point densification module sets a foreground threshold and a background threshold, with the background threshold being greater than the foreground threshold. For a Gaussian point whose semantic label is defined as a foreground region, densification is triggered when the gradient norm of its loss function is greater than the foreground threshold. For Gaussian points whose semantic labels are defined as background regions, densification is triggered when the gradient norm of their loss function is greater than the background threshold.